Enfoque Estadistico del Aprendizaje - Trabajo Final - Regresion Bayesiana

Adrian Marino

Claudio Collado

Inicialización

Fijamos la semilla para poder reproducir los experimentos. También se fija el numero de CPU’s a utilizar.

set.seed(42)
options(mc.cores = 24)

Librerias

Se importan las librerías a utilizar a lo largo de la notebook:

# install.packages(pacman)
# install.packages("https://cran.r-project.org/src/contrib/rstan_2.21.2.tar.gz",repos = NULL,type="source")
# sudo apt-get install libglpk-dev

library(pacman)
p_load(tidyverse, tidymodels, rsample, rstan, shinystan, rstanarm, devtools)

source('../src/dataset.R')
source('../src/plot.R')
source('../src/model.R')

Dataset y Analisis Exploratorio

Palmer Penguins

Palmer Penguins

1. Lectura del dataset

dataset <- load_dataset() %>% mutate_if(is.character, as.factor)

Rows: 344 Columns: 8
── Column specification ───────────────────────────────────────────────────────────────────────────────────
Delimiter: ","
chr (3): species, island, sex
dbl (5): bill_length_mm, bill_depth_mm, flipper_length_mm, body_mass_g, year

ℹ Use `spec()` to retrieve the full column specification for this data.
ℹ Specify the column types or set `show_col_types = FALSE` to quiet this message.

dataset %>% glimpse()

Rows: 344
Columns: 8
$ species           <fct> Adelie, Adelie, Adelie, Adelie, Adelie, Adelie, Adelie, Adelie, Adelie, Adelie,…
$ island            <fct> Torgersen, Torgersen, Torgersen, Torgersen, Torgersen, Torgersen, Torgersen, To…
$ bill_length_mm    <dbl> 39.1, 39.5, 40.3, NA, 36.7, 39.3, 38.9, 39.2, 34.1, 42.0, 37.8, 37.8, 41.1, 38.…
$ bill_depth_mm     <dbl> 18.7, 17.4, 18.0, NA, 19.3, 20.6, 17.8, 19.6, 18.1, 20.2, 17.1, 17.3, 17.6, 21.…
$ flipper_length_mm <dbl> 181, 186, 195, NA, 193, 190, 181, 195, 193, 190, 186, 180, 182, 191, 198, 185, …
$ body_mass_g       <dbl> 3750, 3800, 3250, NA, 3450, 3650, 3625, 4675, 3475, 4250, 3300, 3700, 3200, 380…
$ sex               <fct> male, female, female, NA, female, male, female, male, NA, NA, NA, NA, female, m…
$ year              <dbl> 2007, 2007, 2007, 2007, 2007, 2007, 2007, 2007, 2007, 2007, 2007, 2007, 2007, 2…

2. Variables

Se enumeran y describen breve-mente cada variable que forma parte del dataset:

Variables Numéricas:

• bill_length_mm: Longitud del pico del individuo medida en milímetros (también conocida como longitud del culmen).
• bill_depth_mm: Profundidad del pico del individuo medida en milímetros (también conocida como profundidad del culmen).
• flipper_length_mm: Longitud de la aleta del individuo medida en milímetros.
• body_mass_g: Masa corporal del individuo medida en gramos.
• year: Año en el que se registra el individuo.

Variables Categóricas:

• species: Especie del individuo (Adelie, Gentoo ;) o Chinstrap).
• sex: Sexo del individuo.
• island: Isla donde se encontré el individuo (Biscoe, Dream o Torgersen).

A continuación veamos los posibles valores de las variables categóricas:

show_values(dataset %>% select_if(negate(is.numeric)))

_____________ 
species     n 
============= 
Adelie    152 
Chinstrap  68 
Gentoo    124 
¯¯¯¯¯¯¯¯¯¯¯¯¯ 
_____________ 
island      n 
============= 
Biscoe    168 
Dream     124 
Torgersen  52 
¯¯¯¯¯¯¯¯¯¯¯¯¯ 
__________ 
sex      n 
========== 
female 165 
male   168 
NA      11 
¯¯¯¯¯¯¯¯¯¯ 

3. Resumen de faltantes

missings_summary(dataset)

4. Varibles numericas

hist_plots(dataset)

Observaciones

• Se aprecia que cada año se registro prácticamente el mismo numero de individuos.
• La distribución de la masa corporal de los individuos tiene una asimétrica positiva. Tenemos muchos individuos con valores bajos de masa corporal, con una media de 192 gramos. Luego tenemos menos individuos con valores mas alto.
• La longitud de la aleta parece ser una distribución bi-modal. Tenemos dos modas una 192 mm y otra en 215 mm.
• La longitud del pico también parece tener una ligera simetría positiva. Es decir que lo individuos con menor peso tiene pico mas pequeños.
• Por otro lado la profundidad de pico parece tener una ligera simetría positiva.

box_plots(dataset)

Observaciones

• COMPLETAR

Outliers

No se registran valores mas extremos que el mínimo y máximo valor en cada variables. Es decir que no encontramos outliers.

outliers(dataset, column='bill_length_mm')

$inf
[1] 32.1

$sup
[1] 59.6

outliers(dataset, column='bill_length_mm')

$inf
[1] 32.1

$sup
[1] 59.6

outliers(dataset, column='bill_depth_mm')

$inf
[1] 13.1

$sup
[1] 21.5

outliers(dataset, column='flipper_length_mm')

$inf
[1] 172

$sup
[1] 231

outliers(dataset, column='body_mass_g')

$inf
[1] 2700

$sup
[1] 6300

outliers(dataset, column='year')

$inf
[1] 2007

$sup
[1] 2009

bar_plots(dataset)

Observaciones

• La variable sexo se encuentra balanceada. Por otro lado, contiene algunos valores faltantes.
• La variable island esta completamente desbalanceada. Esto seguramente se debe a una diferencia en numero en las poblaciones en cada isla o a un sesgo al momento de registrar los individuos. Es decir que registramos con individuos en una isla que en otra.
• Lo mismo sucede con las especies de individuos. Vemos un gran desbalance entre la especie Chinstrap vs. otra especies. Por otro aldo Adelie y Gentoo tiene un conteo mas cercano

5. Excluir observaciones con missings

dataset <- dataset %>% drop_na()
missings_summary(dataset)

Warning: attributes are not identical across measure variables;
they will be dropped

6. Correlaciones

corr_plot(dataset %>% dplyr::select(-year))

segmented_pairs_plot(dataset, segment_column='species')

`stat_bin()` using `bins = 30`. Pick better value with `binwidth`.
`stat_bin()` using `bins = 30`. Pick better value with `binwidth`.
`stat_bin()` using `bins = 30`. Pick better value with `binwidth`.
`stat_bin()` using `bins = 30`. Pick better value with `binwidth`.
`stat_bin()` using `bins = 30`. Pick better value with `binwidth`.
`stat_bin()` using `bins = 30`. Pick better value with `binwidth`.
`stat_bin()` using `bins = 30`. Pick better value with `binwidth`.
`stat_bin()` using `bins = 30`. Pick better value with `binwidth`.
`stat_bin()` using `bins = 30`. Pick better value with `binwidth`.
`stat_bin()` using `bins = 30`. Pick better value with `binwidth`.
`stat_bin()` using `bins = 30`. Pick better value with `binwidth`.
`stat_bin()` using `bins = 30`. Pick better value with `binwidth`.
`stat_bin()` using `bins = 30`. Pick better value with `binwidth`.
`stat_bin()` using `bins = 30`. Pick better value with `binwidth`.
`stat_bin()` using `bins = 30`. Pick better value with `binwidth`.

Experimentos

Experimento 1

• Solo variables numéricas.
• Regresión múltiple frecuentista.
• Regresión múltiple bayesiana con priors normales y exponencial.

1. Split train vs. test

train_test <- train_test_split(dataset, train_size = 0.7, shuffle = TRUE)

[1] "Train set size: 233"
[1] "Test set size: 100"

train_set <- train_test[[1]]
test_set  <- train_test[[2]]

2. Modelo lineal

lineal_model_1 <- lm(
  body_mass_g ~ bill_length_mm + bill_depth_mm + flipper_length_mm,
  data = train_set
)

3. Modelo bayesiano

bayesion_model_1 <- stan(
  model_code =  "
    data {
      int<lower=1>               obs_count;
      vector<lower=1>[obs_count] x1;
      vector<lower=1>[obs_count] x2;
      vector<lower=1>[obs_count] x3;
      vector[obs_count]          y;
    }
    parameters {
      real          beta0;
      real          beta1;
      real          beta2;
      real          beta3;
      real<lower=0> sigma;
    }
    model {
      // Distribuciones a priori 
      beta0 ~ normal(0, 8000  /* Fijado por el investigador */ ); // Intercept
      beta1 ~ normal(0, 100   /* Fijado por el investigador */);
      beta2 ~ normal(0, 100   /* Fijado por el investigador */);
      beta3 ~ normal(0, 100   /* Fijado por el investigador */);
      sigma ~ exponential(0.1 /* Fijado por el investigador */);

      // Likelihood    
      y ~ normal(beta0 + beta1 * x1 + beta2 * x2 + beta3 * x3, sigma);
    }
  ",
  data = list(
      obs_count = nrow(train_set),
      y  = colvalues(train_set, 'body_mass_g'),
      x1 = colvalues(train_set, 'bill_length_mm'),
      x2 = colvalues(train_set, 'bill_depth_mm'),
      x3 = colvalues(train_set, 'flipper_length_mm')
  ),
  chains = 3,
  iter   = 300,
  warmup = 180,
  thin   = 1
)

starting worker pid=2744 on localhost:11086 at 19:38:44.884
starting worker pid=2781 on localhost:11086 at 19:38:44.993
starting worker pid=2818 on localhost:11086 at 19:38:45.103

SAMPLING FOR MODEL 'ff7fc5db0fd7100c79e067b72d83d19f' NOW (CHAIN 1).
Chain 1: 
Chain 1: Gradient evaluation took 3.7e-05 seconds
Chain 1: 1000 transitions using 10 leapfrog steps per transition would take 0.37 seconds.
Chain 1: Adjust your expectations accordingly!
Chain 1: 
Chain 1: 
Chain 1: Iteration:   1 / 300 [  0%]  (Warmup)
Chain 1: Iteration:  30 / 300 [ 10%]  (Warmup)
Chain 1: Iteration:  60 / 300 [ 20%]  (Warmup)
Chain 1: Iteration:  90 / 300 [ 30%]  (Warmup)
Chain 1: Iteration: 120 / 300 [ 40%]  (Warmup)
Chain 1: Iteration: 150 / 300 [ 50%]  (Warmup)
Chain 1: Iteration: 180 / 300 [ 60%]  (Warmup)
Chain 1: Iteration: 181 / 300 [ 60%]  (Sampling)
Chain 1: Iteration: 210 / 300 [ 70%]  (Sampling)
Chain 1: Iteration: 240 / 300 [ 80%]  (Sampling)

SAMPLING FOR MODEL 'ff7fc5db0fd7100c79e067b72d83d19f' NOW (CHAIN 2).
Chain 2: 
Chain 2: Gradient evaluation took 4.7e-05 seconds
Chain 2: 1000 transitions using 10 leapfrog steps per transition would take 0.47 seconds.
Chain 2: Adjust your expectations accordingly!
Chain 2: 
Chain 2: 
Chain 2: Iteration:   1 / 300 [  0%]  (Warmup)
Chain 2: Iteration:  30 / 300 [ 10%]  (Warmup)
Chain 1: Iteration: 270 / 300 [ 90%]  (Sampling)
Chain 1: Iteration: 300 / 300 [100%]  (Sampling)
Chain 2: Iteration:  60 / 300 [ 20%]  (Warmup)
Chain 1: 
Chain 1:  Elapsed Time: 0.84136 seconds (Warm-up)
Chain 1:                0.717478 seconds (Sampling)
Chain 1:                1.55884 seconds (Total)
Chain 1: 

SAMPLING FOR MODEL 'ff7fc5db0fd7100c79e067b72d83d19f' NOW (CHAIN 3).
Chain 3: 
Chain 3: Gradient evaluation took 4.4e-05 seconds
Chain 3: 1000 transitions using 10 leapfrog steps per transition would take 0.44 seconds.
Chain 3: Adjust your expectations accordingly!
Chain 3: 
Chain 3: 
Chain 3: Iteration:   1 / 300 [  0%]  (Warmup)
Chain 3: Iteration:  30 / 300 [ 10%]  (Warmup)
Chain 3: Iteration:  60 / 300 [ 20%]  (Warmup)
Chain 2: Iteration:  90 / 300 [ 30%]  (Warmup)
Chain 3: Iteration:  90 / 300 [ 30%]  (Warmup)
Chain 2: Iteration: 120 / 300 [ 40%]  (Warmup)
Chain 3: Iteration: 120 / 300 [ 40%]  (Warmup)
Chain 2: Iteration: 150 / 300 [ 50%]  (Warmup)
Chain 3: Iteration: 150 / 300 [ 50%]  (Warmup)
Chain 2: Iteration: 180 / 300 [ 60%]  (Warmup)
Chain 2: Iteration: 181 / 300 [ 60%]  (Sampling)
Chain 3: Iteration: 180 / 300 [ 60%]  (Warmup)
Chain 3: Iteration: 181 / 300 [ 60%]  (Sampling)
Chain 2: Iteration: 210 / 300 [ 70%]  (Sampling)
Chain 3: Iteration: 210 / 300 [ 70%]  (Sampling)
Chain 2: Iteration: 240 / 300 [ 80%]  (Sampling)
Chain 3: Iteration: 240 / 300 [ 80%]  (Sampling)
Chain 2: Iteration: 270 / 300 [ 90%]  (Sampling)
Chain 3: Iteration: 270 / 300 [ 90%]  (Sampling)
Chain 2: Iteration: 300 / 300 [100%]  (Sampling)
Chain 2: 
Chain 2:  Elapsed Time: 1.16549 seconds (Warm-up)
Chain 2:                0.466704 seconds (Sampling)
Chain 2:                1.63219 seconds (Total)
Chain 2: 
Chain 3: Iteration: 300 / 300 [100%]  (Sampling)
Chain 3: 
Chain 3:  Elapsed Time: 0.802516 seconds (Warm-up)
Chain 3:                0.388908 seconds (Sampling)
Chain 3:                1.19142 seconds (Total)
Chain 3: 

params_1 <- c('beta0', 'beta1', 'beta2', 'beta3', 'sigma')
traceplot(bayesion_model_1, pars = params_1, inc_warmup = TRUE)

4. Coeficientes

lm_vs_br_coeficients(lineal_model_1, bayesion_model_1, params_1)

4. Validación

vars_1 <- c('bill_length_mm', 'bill_depth_mm', 'flipper_length_mm') 

lm_vs_br_models_validation(lineal_model_1, bayesion_model_1, params_1, vars_1, test_set)

bayesion_predictor_1 <- BayesianRegressionPredictor.from(bayesion_model_1, params_1, vars_1)

plot_compare_fit(
  lineal_model_1, 
  bayesion_predictor_1, 
  train_set,
  label_1='Regresion Lineal', 
  label_2='Regresion Bayesiana'
)

Experimento 2

• Idem a experimento 1, incorporando una variable categórica.
• Regresión múltiple frecuentista.
• Regresión bayesiana con priors normales y exponencial.

1. Modelo lineal

lineal_model_2 <- lm(
    body_mass_g 
      ~ bill_length_mm
      + bill_depth_mm
      + flipper_length_mm
      + sex,

  data = train_set
)

2. Modelo bayesiano

Antes que nada transformamos la columna categórica a one-hot encoding:

cat_cols <- c('sex')

train_set_cat <- train_set %>% dummify(cat_cols)
test_set_cat  <- test_set  %>% dummify(cat_cols)

train_set_cat

Construimos una matriz con todas las variables(X) mas el intercept:

to_model_input <- function(df) {
  model.matrix(
    body_mass_g 
      ~ bill_length_mm
      + bill_depth_mm
      + flipper_length_mm
      + sex_female
      + sex_male,

    data = df
  )
}

train_X <- train_set_cat %>% to_model_input()
test_X  <- test_set_cat %>% to_model_input()

Definimos y corremos el modelo bayesiano:

bayesion_model_2 <- stan(
  model_code = "
    data {
      int<lower=1>                 obs_count;
      int<lower=1>                 coef_count;
      matrix[obs_count,coef_count] X;
      vector[obs_count]            y;
    }
    parameters {
      vector[coef_count]  beta;
      real<lower=0>       sigma;
    }
    model {
      // Distribuciones a priori
      beta[1] ~ normal(0, 2000); // Intecept
  
      // Variables numericas
      beta[2] ~ normal(0, 30);
      beta[3] ~ normal(0, 100);
      beta[4] ~ normal(0, 100);
  
      // Sexo
      beta[5] ~ normal(0, 100);
      beta[6] ~ normal(0, 100);
  
      sigma ~ exponential(0.1);
  
      // Likelihood
      y ~ normal(X * beta, sigma);
    }
  ",
  data = list(
      obs_count  = dim(train_X)[1],
      coef_count = dim(train_X)[2],
      y          = colvalues(train_set_cat, 'body_mass_g'),
      X          = train_X
  ),
  chains = 3,
  iter   = 300,
  warmup = 180,
  thin   = 1
)

starting worker pid=3020 on localhost:11086 at 19:39:17.344
starting worker pid=3057 on localhost:11086 at 19:39:17.454
starting worker pid=3094 on localhost:11086 at 19:39:17.563

SAMPLING FOR MODEL '9afeac097877f56e1844b004ab8095d1' NOW (CHAIN 1).
Chain 1: 
Chain 1: Gradient evaluation took 2e-05 seconds
Chain 1: 1000 transitions using 10 leapfrog steps per transition would take 0.2 seconds.
Chain 1: Adjust your expectations accordingly!
Chain 1: 
Chain 1: 
Chain 1: Iteration:   1 / 300 [  0%]  (Warmup)
Chain 1: Iteration:  30 / 300 [ 10%]  (Warmup)
Chain 1: Iteration:  60 / 300 [ 20%]  (Warmup)
Chain 1: Iteration:  90 / 300 [ 30%]  (Warmup)
Chain 1: Iteration: 120 / 300 [ 40%]  (Warmup)
Chain 1: Iteration: 150 / 300 [ 50%]  (Warmup)
Chain 1: Iteration: 180 / 300 [ 60%]  (Warmup)
Chain 1: Iteration: 181 / 300 [ 60%]  (Sampling)
Chain 1: Iteration: 210 / 300 [ 70%]  (Sampling)
Chain 1: Iteration: 240 / 300 [ 80%]  (Sampling)
Chain 1: Iteration: 270 / 300 [ 90%]  (Sampling)
Chain 1: Iteration: 300 / 300 [100%]  (Sampling)
Chain 1: 
Chain 1:  Elapsed Time: 0.422889 seconds (Warm-up)
Chain 1:                0.113815 seconds (Sampling)
Chain 1:                0.536704 seconds (Total)
Chain 1: 

SAMPLING FOR MODEL '9afeac097877f56e1844b004ab8095d1' NOW (CHAIN 2).
Chain 2: 
Chain 2: Gradient evaluation took 2e-05 seconds
Chain 2: 1000 transitions using 10 leapfrog steps per transition would take 0.2 seconds.
Chain 2: Adjust your expectations accordingly!
Chain 2: 
Chain 2: 
Chain 2: Iteration:   1 / 300 [  0%]  (Warmup)
Chain 2: Iteration:  30 / 300 [ 10%]  (Warmup)
Chain 2: Iteration:  60 / 300 [ 20%]  (Warmup)
Chain 2: Iteration:  90 / 300 [ 30%]  (Warmup)
Chain 2: Iteration: 120 / 300 [ 40%]  (Warmup)
Chain 2: Iteration: 150 / 300 [ 50%]  (Warmup)
Chain 2: Iteration: 180 / 300 [ 60%]  (Warmup)
Chain 2: Iteration: 181 / 300 [ 60%]  (Sampling)

SAMPLING FOR MODEL '9afeac097877f56e1844b004ab8095d1' NOW (CHAIN 3).
Chain 3: 
Chain 3: Gradient evaluation took 1.5e-05 seconds
Chain 3: 1000 transitions using 10 leapfrog steps per transition would take 0.15 seconds.
Chain 3: Adjust your expectations accordingly!
Chain 3: 
Chain 3: 
Chain 3: Iteration:   1 / 300 [  0%]  (Warmup)
Chain 3: Iteration:  30 / 300 [ 10%]  (Warmup)
Chain 2: Iteration: 210 / 300 [ 70%]  (Sampling)
Chain 2: Iteration: 240 / 300 [ 80%]  (Sampling)
Chain 2: Iteration: 270 / 300 [ 90%]  (Sampling)
Chain 2: Iteration: 300 / 300 [100%]  (Sampling)
Chain 2: 
Chain 2:  Elapsed Time: 0.499519 seconds (Warm-up)
Chain 2:                0.129509 seconds (Sampling)
Chain 2:                0.629028 seconds (Total)
Chain 2: 
Chain 3: Iteration:  60 / 300 [ 20%]  (Warmup)
Chain 3: Iteration:  90 / 300 [ 30%]  (Warmup)
Chain 3: Iteration: 120 / 300 [ 40%]  (Warmup)
Chain 3: Iteration: 150 / 300 [ 50%]  (Warmup)
Chain 3: Iteration: 180 / 300 [ 60%]  (Warmup)
Chain 3: Iteration: 181 / 300 [ 60%]  (Sampling)
Chain 3: Iteration: 210 / 300 [ 70%]  (Sampling)
Chain 3: Iteration: 240 / 300 [ 80%]  (Sampling)
Chain 3: Iteration: 270 / 300 [ 90%]  (Sampling)
Chain 3: Iteration: 300 / 300 [100%]  (Sampling)
Chain 3: 
Chain 3:  Elapsed Time: 0.498432 seconds (Warm-up)
Chain 3:                0.1909 seconds (Sampling)
Chain 3:                0.689332 seconds (Total)
Chain 3: 

params_2 <- c('beta[1]', 'beta[2]', 'beta[3]', 'beta[4]', 'beta[5]', 'beta[6]', 'sigma')
traceplot(bayesion_model_2, inc_warmup = TRUE, pars = params_2)

3. Coeficientes

lineal_model_2$coefficients

      (Intercept)    bill_length_mm     bill_depth_mm flipper_length_mm           sexmale 
     -2399.959627         -1.251644        -81.390712         38.730138        509.703109 

br_coeficients(bayesion_model_2, params_2)

4. Validación

vars_2 <- c('bill_length_mm', 'bill_depth_mm', 'flipper_length_mm', 'sex_female','sex_male')

lm_vs_br_models_validation(
  lineal_model_2, 
  bayesion_model_2, 
  params_2,
  vars_2,
  test_set, 
  test_set_cat
)

bayesion_predictor_2 <- BayesianRegressionPredictor.from(bayesion_model_2, params_2, vars_2)

plot_compare_fit(
  lineal_model_2, 
  bayesion_predictor_2, 
  test_set, 
  test_set_cat,
  label_1='Regresion Lineal', 
  label_2='Regresion Bayesiana'
)

Experimento 3

• Idem a experimento 1 incorporando outliers en alguna variable numérica.
• Regresión múltiple frecuentista.
• Regresión bayesiana con priors normales y exponencial.

1. Outliers

A continuación vamos a agregar outliers a la variable flipper_length_mm, la cual define la longitud de la aleta del individuo medida en milímetros.

Para visualizar los nuevos outliers a continuación graficamos flipper_length_mm vs. body_mass_g antes y después de agregar outliers:

plot_data(train_set)

train_set_with_outliers <- train_set %>%
  mutate(flipper_length_mm = ifelse(
    body_mass_g > 4900 & body_mass_g < 5000, 
    flipper_length_mm + (flipper_length_mm * runif(1, 0.1, 0.2)), 
    flipper_length_mm
  ))

plot_data(train_set_with_outliers)

2. Modelo lineal

lineal_model_3 <- lm(
  body_mass_g ~ bill_length_mm + bill_depth_mm + flipper_length_mm,
  data = train_set_with_outliers
)

Comparemos el ajuste del modelo lineal ajustando en un dataset de train con y sin outliers:

plot_compare_fit(
  lineal_model_1, 
  lineal_model_3, 
  train_set_with_outliers,
  label_1='Regresión Lineal SIN outliers', 
  label_2='Regresión Lineal CON outliters'
)

Se aprecia que el modelo entrenado en el train set outliers esta apalancado por las observaciones atipicas.

2. Modelo lineal Robusto

Entrenamos una regresión lineal múltiple robusta para intentar de disminuir el efecto de los nuevos outliers.

p_load(MASS)
robust_lineal_model_3 <- rlm(
  body_mass_g ~ bill_length_mm + bill_depth_mm + flipper_length_mm,
  data = train_set_with_outliers
)

plot_compare_fit(
  lineal_model_1, 
  robust_lineal_model_3, 
  train_set_with_outliers,
  label_1='Regresión Lineal SIN outliers', 
  label_2='Regresión Lineal Robusta CON outliters'
)

Gráficamente no se llega a distinguir pero el modelo robusto termina ajustando mejor que el modelo lineal clásico. Esto se puede observar cuando comparamos el RMSE/MAE sobre train.

lm_vs_lm_models_validation(lineal_model_3, robust_lineal_model_3, train_set_with_outliers)

Warning in actual - predicted :
  longer object length is not a multiple of shorter object length
Warning in actual - predicted :
  longer object length is not a multiple of shorter object length
Warning in actual - predicted :
  longer object length is not a multiple of shorter object length
Warning in actual - predicted :
  longer object length is not a multiple of shorter object length

Mas alla de esto el modelo clasico sigue dando mejores resultado en test:

lm_vs_lm_models_validation(lineal_model_3, robust_lineal_model_3, test_set)

3. Modelo bayesiano con Likelihood normal

bayesion_model_3 <- stan(
  model_code =  "
    data {
      int<lower=1>               obs_count;
      vector<lower=1>[obs_count] x1;
      vector<lower=1>[obs_count] x2;
      vector<lower=1>[obs_count] x3;
      vector[obs_count]          y;
    }
    parameters {
      real          beta0;
      real          beta1;
      real          beta2;
      real          beta3;
      real<lower=0> sigma;
    }
    model {
      // Distribuciones a priori
      beta0 ~ normal(0, 8000); // Intecept
      beta1 ~ normal(0, 100);
      beta2 ~ normal(0, 100);
      beta3 ~ normal(0, 100);
      sigma ~ exponential(0.1);
    
      // Likelihood    
      y ~ normal(beta0 + beta1 * x1 + beta2 * x2 + beta3 * x3, sigma);
    }
  ",
  data = list(
      obs_count = nrow(train_set_with_outliers),
      y  = colvalues(train_set_with_outliers, 'body_mass_g'),
      x1 = colvalues(train_set_with_outliers, 'bill_length_mm'),
      x2 = colvalues(train_set_with_outliers, 'bill_depth_mm'),
      x3 = colvalues(train_set_with_outliers, 'flipper_length_mm')
  ),
  chains = 3,
  iter   = 300,
  warmup = 180,
  thin   = 1
)

starting worker pid=3296 on localhost:11086 at 19:39:48.793
starting worker pid=3333 on localhost:11086 at 19:39:48.903
starting worker pid=3370 on localhost:11086 at 19:39:49.012

SAMPLING FOR MODEL 'c0c5a3ba8590ba28e2cb449cd35b7ade' NOW (CHAIN 1).
Chain 1: 
Chain 1: Gradient evaluation took 3.3e-05 seconds
Chain 1: 1000 transitions using 10 leapfrog steps per transition would take 0.33 seconds.
Chain 1: Adjust your expectations accordingly!
Chain 1: 
Chain 1: 
Chain 1: Iteration:   1 / 300 [  0%]  (Warmup)
Chain 1: Iteration:  30 / 300 [ 10%]  (Warmup)
Chain 1: Iteration:  60 / 300 [ 20%]  (Warmup)
Chain 1: Iteration:  90 / 300 [ 30%]  (Warmup)
Chain 1: Iteration: 120 / 300 [ 40%]  (Warmup)

SAMPLING FOR MODEL 'c0c5a3ba8590ba28e2cb449cd35b7ade' NOW (CHAIN 2).
Chain 2: 
Chain 2: Gradient evaluation took 4.7e-05 seconds
Chain 2: 1000 transitions using 10 leapfrog steps per transition would take 0.47 seconds.
Chain 2: Adjust your expectations accordingly!
Chain 2: 
Chain 2: 
Chain 2: Iteration:   1 / 300 [  0%]  (Warmup)
Chain 2: Iteration:  30 / 300 [ 10%]  (Warmup)
Chain 1: Iteration: 150 / 300 [ 50%]  (Warmup)
Chain 1: Iteration: 180 / 300 [ 60%]  (Warmup)
Chain 1: Iteration: 181 / 300 [ 60%]  (Sampling)
Chain 2: Iteration:  60 / 300 [ 20%]  (Warmup)
Chain 1: Iteration: 210 / 300 [ 70%]  (Sampling)
Chain 1: Iteration: 240 / 300 [ 80%]  (Sampling)

SAMPLING FOR MODEL 'c0c5a3ba8590ba28e2cb449cd35b7ade' NOW (CHAIN 3).
Chain 3: 
Chain 3: Gradient evaluation took 4.3e-05 seconds
Chain 3: 1000 transitions using 10 leapfrog steps per transition would take 0.43 seconds.
Chain 3: Adjust your expectations accordingly!
Chain 3: 
Chain 3: 
Chain 3: Iteration:   1 / 300 [  0%]  (Warmup)
Chain 3: Iteration:  30 / 300 [ 10%]  (Warmup)
Chain 1: Iteration: 270 / 300 [ 90%]  (Sampling)
Chain 1: Iteration: 300 / 300 [100%]  (Sampling)
Chain 1: 
Chain 1:  Elapsed Time: 1.31268 seconds (Warm-up)
Chain 1:                0.434985 seconds (Sampling)
Chain 1:                1.74766 seconds (Total)
Chain 1: 
Chain 2: Iteration:  90 / 300 [ 30%]  (Warmup)
Chain 3: Iteration:  60 / 300 [ 20%]  (Warmup)
Chain 2: Iteration: 120 / 300 [ 40%]  (Warmup)
Chain 2: Iteration: 150 / 300 [ 50%]  (Warmup)
Chain 2: Iteration: 180 / 300 [ 60%]  (Warmup)
Chain 2: Iteration: 181 / 300 [ 60%]  (Sampling)
Chain 3: Iteration:  90 / 300 [ 30%]  (Warmup)
Chain 2: Iteration: 210 / 300 [ 70%]  (Sampling)
Chain 2: Iteration: 240 / 300 [ 80%]  (Sampling)
Chain 3: Iteration: 120 / 300 [ 40%]  (Warmup)
Chain 2: Iteration: 270 / 300 [ 90%]  (Sampling)
Chain 3: Iteration: 150 / 300 [ 50%]  (Warmup)
Chain 2: Iteration: 300 / 300 [100%]  (Sampling)
Chain 2: 
Chain 2:  Elapsed Time: 1.1051 seconds (Warm-up)
Chain 2:                0.495396 seconds (Sampling)
Chain 2:                1.6005 seconds (Total)
Chain 2: 
Chain 3: Iteration: 180 / 300 [ 60%]  (Warmup)
Chain 3: Iteration: 181 / 300 [ 60%]  (Sampling)
Chain 3: Iteration: 210 / 300 [ 70%]  (Sampling)
Chain 3: Iteration: 240 / 300 [ 80%]  (Sampling)
Chain 3: Iteration: 270 / 300 [ 90%]  (Sampling)
Chain 3: Iteration: 300 / 300 [100%]  (Sampling)
Chain 3: 
Chain 3:  Elapsed Time: 1.25576 seconds (Warm-up)
Chain 3:                0.638169 seconds (Sampling)
Chain 3:                1.89393 seconds (Total)
Chain 3: 

params_3 <- c('beta0', 'beta1', 'beta2', 'beta3', 'sigma')
traceplot(bayesion_model_3, pars = params_3, inc_warmup = TRUE)

4. Coeficientes

lm_vs_br_coeficients(robust_lineal_model_3, bayesion_model_3, params_3)

5. Validación

vars_3 <- c('bill_length_mm', 'bill_depth_mm', 'flipper_length_mm') 

lm_vs_br_models_validation(robust_lineal_model_3, bayesion_model_3, params_3, vars_3, test_set)

bayesion_predictor_3 <- BayesianRegressionPredictor.from(bayesion_model_3, params_3, vars_3)

plot_compare_fit(
  robust_lineal_model_3, 
  bayesion_predictor_3, 
  train_set,
  label_1='Regresion Lineal Robusta CON outliers', 
  label_2='Regresion Bayesiana CON outliers'
)

6. Modelo bayesiano con Likelihood t-student

COMPLETAR

Experimento 4

• Idem experimento 1 pero reduciendo la cantidad de observaciones a pocos valores (ej:30).
• Regresion multiple frecuentista.
• Regresion bayesiana con priors normales y exponencial.

1. Split train - test

En este aso entrenamos solo con el 10% de lo datos.

train_test <- train_test_split(dataset, train_size = 0.05, shuffle = TRUE)

[1] "Train set size: 16"
[1] "Test set size: 317"

train_set_4 <- train_test[[1]]
test_set_4  <- train_test[[2]]

plot_data(train_set_4)

2. Modelo lineal

lineal_model_4 <- lm(
  body_mass_g ~ bill_length_mm + bill_depth_mm + flipper_length_mm,
  data = train_set_4
)

3. Modelo bayesiano

bayesion_model_4 <- stan(
  model_code =  "
    data {
      int<lower=1>               obs_count;
      vector<lower=1>[obs_count] x1;
      vector<lower=1>[obs_count] x2;
      vector<lower=1>[obs_count] x3;
      vector[obs_count]          y;
    }
    parameters {
      real          beta0;
      real          beta1;
      real          beta2;
      real          beta3;
      real<lower=0> sigma;
    }
    model {
      // Distribuciones a priori
      beta0 ~ normal(0, 8000); // Intercept
      beta1 ~ normal(0, 100);
      beta2 ~ normal(0, 100);
      beta3 ~ normal(0, 100);
      sigma ~ exponential(0.1);
    
      // Likelihood
      y ~ normal(beta0 + beta1 * x1 + beta2 * x2 + beta3 * x3, sigma);
    }
  ",
  data = list(
      obs_count = nrow(train_set_4),
      y  = colvalues(train_set_4, 'body_mass_g'),
      x1 = colvalues(train_set_4, 'bill_length_mm'),
      x2 = colvalues(train_set_4, 'bill_depth_mm'),
      x3 = colvalues(train_set_4, 'flipper_length_mm')
  ),
  chains = 3,
  iter   = 300,
  warmup = 180,
  thin   = 1
)

starting worker pid=3572 on localhost:11086 at 19:40:19.482
starting worker pid=3609 on localhost:11086 at 19:40:19.591
starting worker pid=3646 on localhost:11086 at 19:40:19.700

SAMPLING FOR MODEL 'a680c5a9703a67be8bd5c21746d8ddcd' NOW (CHAIN 1).
Chain 1: 
Chain 1: Gradient evaluation took 1.5e-05 seconds
Chain 1: 1000 transitions using 10 leapfrog steps per transition would take 0.15 seconds.
Chain 1: Adjust your expectations accordingly!
Chain 1: 
Chain 1: 
Chain 1: Iteration:   1 / 300 [  0%]  (Warmup)
Chain 1: Iteration:  30 / 300 [ 10%]  (Warmup)
Chain 1: Iteration:  60 / 300 [ 20%]  (Warmup)
Chain 1: Iteration:  90 / 300 [ 30%]  (Warmup)
Chain 1: Iteration: 120 / 300 [ 40%]  (Warmup)
Chain 1: Iteration: 150 / 300 [ 50%]  (Warmup)
Chain 1: Iteration: 180 / 300 [ 60%]  (Warmup)
Chain 1: Iteration: 181 / 300 [ 60%]  (Sampling)
Chain 1: Iteration: 210 / 300 [ 70%]  (Sampling)
Chain 1: Iteration: 240 / 300 [ 80%]  (Sampling)
Chain 1: Iteration: 270 / 300 [ 90%]  (Sampling)
Chain 1: Iteration: 300 / 300 [100%]  (Sampling)
Chain 1: 
Chain 1:  Elapsed Time: 0.154591 seconds (Warm-up)
Chain 1:                0.046818 seconds (Sampling)
Chain 1:                0.201409 seconds (Total)
Chain 1: 

SAMPLING FOR MODEL 'a680c5a9703a67be8bd5c21746d8ddcd' NOW (CHAIN 2).
Chain 2: 
Chain 2: Gradient evaluation took 1.2e-05 seconds
Chain 2: 1000 transitions using 10 leapfrog steps per transition would take 0.12 seconds.
Chain 2: Adjust your expectations accordingly!
Chain 2: 
Chain 2: 
Chain 2: Iteration:   1 / 300 [  0%]  (Warmup)
Chain 2: Iteration:  30 / 300 [ 10%]  (Warmup)
Chain 2: Iteration:  60 / 300 [ 20%]  (Warmup)
Chain 2: Iteration:  90 / 300 [ 30%]  (Warmup)
Chain 2: Iteration: 120 / 300 [ 40%]  (Warmup)
Chain 2: Iteration: 150 / 300 [ 50%]  (Warmup)
Chain 2: Iteration: 180 / 300 [ 60%]  (Warmup)
Chain 2: Iteration: 181 / 300 [ 60%]  (Sampling)
Chain 2: Iteration: 210 / 300 [ 70%]  (Sampling)
Chain 2: Iteration: 240 / 300 [ 80%]  (Sampling)
Chain 2: Iteration: 270 / 300 [ 90%]  (Sampling)
Chain 2: Iteration: 300 / 300 [100%]  (Sampling)
Chain 2: 
Chain 2:  Elapsed Time: 0.113693 seconds (Warm-up)
Chain 2:                0.063825 seconds (Sampling)
Chain 2:                0.177518 seconds (Total)
Chain 2: 

SAMPLING FOR MODEL 'a680c5a9703a67be8bd5c21746d8ddcd' NOW (CHAIN 3).
Chain 3: 
Chain 3: Gradient evaluation took 1.1e-05 seconds
Chain 3: 1000 transitions using 10 leapfrog steps per transition would take 0.11 seconds.
Chain 3: Adjust your expectations accordingly!
Chain 3: 
Chain 3: 
Chain 3: Iteration:   1 / 300 [  0%]  (Warmup)
Chain 3: Iteration:  30 / 300 [ 10%]  (Warmup)
Chain 3: Iteration:  60 / 300 [ 20%]  (Warmup)
Chain 3: Iteration:  90 / 300 [ 30%]  (Warmup)
Chain 3: Iteration: 120 / 300 [ 40%]  (Warmup)
Chain 3: Iteration: 150 / 300 [ 50%]  (Warmup)
Chain 3: Iteration: 180 / 300 [ 60%]  (Warmup)
Chain 3: Iteration: 181 / 300 [ 60%]  (Sampling)
Chain 3: Iteration: 210 / 300 [ 70%]  (Sampling)
Chain 3: Iteration: 240 / 300 [ 80%]  (Sampling)
Chain 3: Iteration: 270 / 300 [ 90%]  (Sampling)
Chain 3: Iteration: 300 / 300 [100%]  (Sampling)
Chain 3: 
Chain 3:  Elapsed Time: 0.126041 seconds (Warm-up)
Chain 3:                0.024386 seconds (Sampling)
Chain 3:                0.150427 seconds (Total)
Chain 3: 

params_4 <- c('beta0', 'beta1', 'beta2', 'beta3', 'sigma')
traceplot(bayesion_model_4, pars = params_4, inc_warmup = TRUE)

4. Coeficientes

Coeficientes de la regresión múltiple:

lineal_model_4$coefficients

      (Intercept)    bill_length_mm     bill_depth_mm flipper_length_mm 
      -8874.92522          24.46866          76.75274          53.51397 

Coeficientes descubiertos por la regresión múltiple bayesiana:

for(param in params_4) print(get_posterior_mean(bayesion_model_4, par=param)[4])

[1] -8223.236
[1] 23.60412
[1] 61.40233
[1] 51.73701
[1] 241.8201

5. Validación

vars_4 <- c('bill_length_mm', 'bill_depth_mm', 'flipper_length_mm') 

lm_vs_br_models_validation(lineal_model_4, bayesion_model_4, params_4, vars_4, test_set)

bayesion_predictor_4 <- BayesianRegressionPredictor.from(bayesion_model_4, params_4, vars_4)

plot_compare_fit(
  lineal_model_4,
  bayesion_predictor_4, 
  train_set,
  label_1='Regresion Lineal', 
  label_2='Regresion Bayesiana'
)

Experimento 5

• Igual al experimento 1 pero proponiendo dos nuevas regresiones bayesianas con priors para los parámetros que sean:
  • Una poca informativa (uniforme).
  • Una muy informativa (sesgada o con muy poca varianza).
• Comparar con resultados de la bayesiana del experimento A

1. Modelo bayesiano con parametro con distribucion poco informativa

1. Modelo

Definimos una distribución uniforme para el beta asociado a la variable flipper_length_mm.

starting worker pid=3848 on localhost:11086 at 19:40:49.175
starting worker pid=3885 on localhost:11086 at 19:40:49.287
starting worker pid=3922 on localhost:11086 at 19:40:49.396

SAMPLING FOR MODEL 'ec5409959af1c66bd72a3906ea8aa743' NOW (CHAIN 1).
Chain 1: 
Chain 1: Gradient evaluation took 3.6e-05 seconds
Chain 1: 1000 transitions using 10 leapfrog steps per transition would take 0.36 seconds.
Chain 1: Adjust your expectations accordingly!
Chain 1: 
Chain 1: 
Chain 1: Iteration:   1 / 1000 [  0%]  (Warmup)
Chain 1: Iteration: 100 / 1000 [ 10%]  (Warmup)
Chain 1: Iteration: 181 / 1000 [ 18%]  (Sampling)

SAMPLING FOR MODEL 'ec5409959af1c66bd72a3906ea8aa743' NOW (CHAIN 2).
Chain 2: 
Chain 2: Gradient evaluation took 4.4e-05 seconds
Chain 2: 1000 transitions using 10 leapfrog steps per transition would take 0.44 seconds.
Chain 2: Adjust your expectations accordingly!
Chain 2: 
Chain 2: 
Chain 2: Iteration:   1 / 1000 [  0%]  (Warmup)
Chain 1: Iteration: 280 / 1000 [ 28%]  (Sampling)
Chain 2: Iteration: 100 / 1000 [ 10%]  (Warmup)

SAMPLING FOR MODEL 'ec5409959af1c66bd72a3906ea8aa743' NOW (CHAIN 3).
Chain 3: Rejecting initial value:
Chain 3:   Error evaluating the log probability at the initial value.
Chain 3: Exception: exponential_lpdf: Random variable is -0.274745, but must be >= 0!  (in 'model19f65fff8b1_ec5409959af1c66bd72a3906ea8aa743' at line 21)

Chain 3: 
Chain 3: Gradient evaluation took 4e-05 seconds
Chain 3: 1000 transitions using 10 leapfrog steps per transition would take 0.4 seconds.
Chain 3: Adjust your expectations accordingly!
Chain 3: 
Chain 3: 
Chain 3: Iteration:   1 / 1000 [  0%]  (Warmup)
Chain 1: Iteration: 380 / 1000 [ 38%]  (Sampling)
Chain 2: Iteration: 181 / 1000 [ 18%]  (Sampling)
Chain 2: Iteration: 280 / 1000 [ 28%]  (Sampling)
Chain 1: Iteration: 480 / 1000 [ 48%]  (Sampling)
Chain 2: Iteration: 380 / 1000 [ 38%]  (Sampling)
Chain 3: Iteration: 100 / 1000 [ 10%]  (Warmup)
Chain 2: Iteration: 480 / 1000 [ 48%]  (Sampling)
Chain 1: Iteration: 580 / 1000 [ 58%]  (Sampling)
Chain 3: Iteration: 181 / 1000 [ 18%]  (Sampling)
Chain 2: Iteration: 580 / 1000 [ 58%]  (Sampling)
Chain 2: Iteration: 680 / 1000 [ 68%]  (Sampling)
Chain 1: Iteration: 680 / 1000 [ 68%]  (Sampling)
Chain 3: Iteration: 280 / 1000 [ 28%]  (Sampling)
Chain 2: Iteration: 780 / 1000 [ 78%]  (Sampling)
Chain 2: Iteration: 880 / 1000 [ 88%]  (Sampling)
Chain 1: Iteration: 780 / 1000 [ 78%]  (Sampling)
Chain 3: Iteration: 380 / 1000 [ 38%]  (Sampling)
Chain 2: Iteration: 980 / 1000 [ 98%]  (Sampling)
Chain 2: Iteration: 1000 / 1000 [100%]  (Sampling)
Chain 2: 
Chain 2:  Elapsed Time: 1.01022 seconds (Warm-up)
Chain 2:                2.02517 seconds (Sampling)
Chain 2:                3.0354 seconds (Total)
Chain 2: 
Chain 1: Iteration: 880 / 1000 [ 88%]  (Sampling)
Chain 3: Iteration: 480 / 1000 [ 48%]  (Sampling)
Chain 1: Iteration: 980 / 1000 [ 98%]  (Sampling)
Chain 3: Iteration: 580 / 1000 [ 58%]  (Sampling)
Chain 1: Iteration: 1000 / 1000 [100%]  (Sampling)
Chain 1: 
Chain 1:  Elapsed Time: 1.10362 seconds (Warm-up)
Chain 1:                3.90271 seconds (Sampling)
Chain 1:                5.00633 seconds (Total)
Chain 1: 
Chain 3: Iteration: 680 / 1000 [ 68%]  (Sampling)
Chain 3: Iteration: 780 / 1000 [ 78%]  (Sampling)
Chain 3: Iteration: 880 / 1000 [ 88%]  (Sampling)
Chain 3: Iteration: 980 / 1000 [ 98%]  (Sampling)
Chain 3: Iteration: 1000 / 1000 [100%]  (Sampling)
Chain 3: 
Chain 3:  Elapsed Time: 1.14468 seconds (Warm-up)
Chain 3:                3.67208 seconds (Sampling)
Chain 3:                4.81675 seconds (Total)
Chain 3: 

2. Coeficientes

br_vs_br_coeficients(bayesion_model_1, bayesion_model_5, params_5)

3. Validación

lm_vs_br_models_validation(lineal_model_1, bayesion_model_1, params_1, vars_1, test_set)

4. Validacion

vars_5 <- c('bill_length_mm', 'bill_depth_mm', 'flipper_length_mm') 

lm_vs_br_models_validation(lineal_model_1, bayesion_model_5, params_5, vars_5, test_set)

bayesion_predictor_5 <- BayesianRegressionPredictor.from(bayesion_model_5, params_5, vars_5)

plot_compare_fit(
  bayesion_predictor_1,
  bayesion_predictor_5,
  train_set,
  label_1='Regresion Bayesiana con dist informativa', 
  label_2='Regresion Bayesiana con dist menos informativa'
)

2. Modelo bayesiano con parametro con distribucion muy informativa sesgada o con poca varianza

COMPLETAR

Referencias

• Making Predictions from Stan models in R
• How to represent a categorical predictor rstan?
• R commons